Nabin K. Malakar, Ph.D.

I am a computational physicist working on societal applications of machine-learning techniques.

Research Links

My research interests span multi-disciplinary fields involving Societal applications of Machine Learning, Decision-theoretic approach to automated Experimental Design, Bayesian statistical data analysis and signal processing.


Interested about the picture? Autonomous experimental design allows us to answer the question of where to take the measurements. More about it is here...


I addition to the research, I also like to hike, bike, read and play with water color.

Thanks for the visit. Please feel free to visit my Weblogs.

Welcome to Please visit again.

Tuesday, December 30, 2014

#AMS2015, January 04 - 08, 2015 Phoenix, AZ #conference @ametsoc

Data fusion of Satellite AOD and WRF meteorology for improved PM25 estimation for northeast USA

Monday, 5 January 2015: 1:45 PM 
at Sixth Conference on Environment and Health)
228AB (Phoenix Convention Center - West and North Buildings)
Nabin Malakar, City College of New York, New York, NY; and L. Cordero, B. Gross, D. Vidal, and F. Moshary
The current approach to ingesting satellite data (IDEA- Infusing satellite Data into Environmental air quality Applications Product) into surface PM2.5 retrievals uses a combination of spatial interpolation and a global geo-chemical model (GEOS-CHEM) to define appropriate mass to AOD factor maps that can be used with satellite AOD retreivals. This information is then statistically blended with current AIRNow measurements creating a refined retrieval product. In this paper, we propose to use the same approach except that we replace the GEOS-CHEM component with an alternative high resolution meteorological model scheme. In particular, we illustrate that the GEOS-CHEM factors can be strongly biased and explore methods that incorporate a combination of satellite AOD retrievals with WRF meteorological forecasts on a regional scale. We find that although PBL height should be a significant factor, the WRF model uncertainties for PBL height in comparison to Calipso make this factor less reliable. More directly we find that the covarying PBL averaged temperature (together with wind direction) are the most important factors. Direct statistical comparisons are made against the IDEA product showing the utility of this approach over regional scales. In addition, we explore the importance of a number of factors including season and time averaging showing that the satellite approach improves significantly as the time averaging window decreases illustrating the potential impact that GOES-R will have on PM25 estimation.

Fusing Spatial Kriging with Satellite Estimates to Obtain a Regional Estimation of PM2.5

Daniel Vidal, City College of New York, New York, NY; and B. Gross, N. Malakar, and L. Cordero
This work focuses on developing estimates of ground-level fine particulate matter (PM2.5) in the northeastern U.S. based on measurements derived from the Air Quality System (AQS) repository. Real time monitoring of PM2.5 is important due to its effect on climate change and human health, however, designated samplers used by state agencies do not provide optimal spatial coverage given their high cost and extensive human labor dependence. Through the application of remote sensing instruments, information about PM2.5 concentrations can be generated at certain locations. On the other hand, coverage limitation also occurs when using satellite remote sensing methods due to atmospheric conditions. Therefore, our approach begins by utilizing surface PM2.5 measurements collected from the Remote Sensing Information Gateway (RSIG) portal in order to build fine particulate matter estimations by applying a Spatial Kriging technique. Then, we combine our Kriging estimations to the satellite derived PM2.5 obtained through an Artificial Neural Network (ANN) scheme to generate a daily regional PM2.5 product. Finally, evaluation of our fused algorithm's technique is assessed by performing comparisons against Kriging and neural network individual performances, showing the promising value added by the combination of these two techniques in producing more accurate estimations of surface level PM2.5 over our region of interest.

This one is related to the award winning work by Daniel:

Analysis of New York City traffic data, land use, emissions and high resolution local meteorology for the prediction of neighborhood scale intra-urban PM2.5 and O3
Monday, 5 January 2015: 4:30 PM 

at Sixth Conference on Environment and Health)
228AB (Phoenix Convention Center - West and North Buildings)
Chowdhary Nazmi, NOAA/CREST/City College, New York, NY; and N. Malakar, L. Cordero, and B. Gross
Air pollution affects the health and well-being of residents of mega cities like New York. Predicting the air pollutant concentration throughout the city can be difficult because the sources and levels of the pollutants can vary from season to season. Local meteorology, traffic and land use also play an important role in these variations and the use of statistical machine learning tools such as Neural Networks can be very useful. In order to develop a Neural Network for the prediction of intra-urban air pollutants (PM2.5, O3), high resolution local data are collected and analyzed. Surface level high resolution temperature, relative humidity and wind speed data are collected from the CCNY METNET network. Annual average daily traffic data from NYMTC model as well as continuous and short count traffic data are collected from NYSDOT. High density data from NYC Community Air Survey model is used to analyze the relationship between background and street level indicators for PM2.5 and O3. All the variables (meteorology, population, traffic, land use etc) are ranked according to the absolute strength of their correlation with the measured pollutants and highest ranking variables are identified to be used for the development of a Neural Network. An analysis of how street level pollution differs from background AIRNow observations will be made showing the importance of high density observations. The potential to use the model in other urban areas will also be explored.

Having now relocated to NASA JPL, it is fun to reflect back to see what was accomplished during my stay at CCNY.

Friday, December 19, 2014

Presented in the AGU 2014, San Francisco, CA

  • GC51D-0460Ingesting Land Surface Temperature differences to improve Downwelling Solar Radiation using Artificial Neural Network: A Case Study
  • In order to study the effects of global climate change on regional scales, we need high resolution models that can be injected into local ecosystem models. Although the injection of regional Meteorological Models such as Weather Research and Forecasting (WRF) can be attempted where the Global Circulation Model (GCM) conditions and the forecasted land surface properties are encoded into future time slices - this approach is extremely computer intensive.
    We present a two-step mechanism in which low resolution meteorological data including both surface and column integrated parameters are combined with high resolution land surface classification parameters to improve on purely interpolative approaches by using machine learning techniques. In particular, we explore the improvement of surface radiation estimates critical for ecosystem modeling by combining both model and satellite based surface radiation together with land surface temperature differences.

    Nabin Malakar - NASA Jet Propulsion Laboratory
    Mark Bailey
    CUNY City College
    Rebecca Latto
    CUNY City College
    Emmanuel Ekwedike
    CUNY City College
    Barry Gross
    CUNY City College
    Jorge Gonzalez
    CUNY City College
    Charles Vorosmarty
    CUNY City College
    Glynn Hulley - NASA Jet Propulsion Laboratory

    A51B-3024Bias Correction of MODIS AOD using DragonNET to obtain improved estimation of PM2.5

MODIS AOD retreivals using the Dark Target algorithm is strongly affected by the underlying surface reflection properties. In particular, the operational algorithms make use of surface parameterizations trained on global datasets and therefore do not account properly for urban surface differences. This parameterization continues to show an underestimation of the surface reflection which results in a general over-biasing in AOD retrievals. Recent results using the Dragon-Network datasets as well as high resolution retrievals in the NYC area illustrate that this is even more significant at the newest C006 3 km retrievals. In the past, we used AERONET observation in the City College to obtain bias-corrected AOD, but the homogeneity assumptions using only one site for the region is clearly an issue. On the other hand, DragonNET observations provide ample opportunities to obtain better tuning the surface corrections while also providing better statistical validation. In this study we present a neural network method to obtain bias correction of the MODIS AOD using multiple factors including surface reflectivity at 2130nm, sun-view geometrical factors and land-class information. These corrected AOD’s are then used together with additional WRF meteorological factors to improve estimates of PM2.5. Efforts to explore the portability to other urban areas will be discussed. In addition, annual surface ratio maps will be developed illustrating that among the land classes, the urban pixels constitute the largest deviations from the operational model.